Method and system for unmanned aerial vehicle and unmanned vehicle to cooperatively identify camouflaged targets

Abstract

The application provides a method and system for detecting camouflaged targets by unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs). The system ensures the synchronization of UAVs and UGVs through the global positioning system (GPS), and uses the infrared cameras, binocular cameras, and radar sensors carried by both to efficiently collect data in the target area. The system further preprocesses and analyzes the collected data, and uses improved SSD algorithms and pattern recognition techniques to identify potential camouflaged targets, and performs pattern matching and information fusion to improve the accuracy of the identification results. The SE attention module is introduced innovatively, which optimizes the feature fusion process, enhances the network's spatial correlation capture ability, and enables the system to use global information more effectively, improving the detection accuracy of camouflaged targets. The application solves the problem of poor reliability in traditional methods for detecting camouflaged targets, improves the real-time performance and effectiveness of the cooperative operation of UAVs and UGVs, and significantly improves the accuracy and response speed of camouflaged target identification in battlefield environments. It has important theoretical value and application prospect in various military crisis response.

Description

Technical Field

[0001] This patent relates to computer vision and artificial intelligence, and in particular to methods and systems for collaborative identification of camouflaged targets by drones and unmanned vehicles. Background Technology

[0002] In military tactics, camouflage refers to the process of disguising military targets or resources to blend their appearance into the surrounding environment, making them visually difficult to detect and identify. This camouflage can include the use of camouflage paint, bunkers, and simulated terrain, aiming to reduce the enemy's ability to detect and engage the target, thereby increasing the success rate and security of military operations.

[0003] Currently, the detection of camouflaged targets is of great significance in the military field. The identification of camouflaged targets is a complex task handled through machine learning, particularly deep learning methods. These include using convolutional neural networks (CNNs) to automatically process and learn data features, and Region Proposal Networks (R-CNN) and its variants such as Fast R-CNN and Faster R-CNN to effectively improve recognition efficiency by generating candidate regions that may contain targets. In addition, Generative Adversarial Networks (GANs) can learn and improve the distinguishability between camouflaged targets and the background, and attention mechanisms can mimic how humans focus on visual scenes, weighting key regions to improve recognition accuracy. Furthermore, multi-task learning and multi-scale, multi-view analysis also play a crucial role in providing comprehensive features and improving recognition capabilities.

[0004] However, these techniques often rely on manually designed features such as edge detection, texture analysis, and color histograms. While these features may be effective in specific scenarios, they are insufficient to handle the full complexity of camouflaged targets, especially when the background and target are at similar heights. Background interference is also a major problem; traditional detection methods may fail in noisy backgrounds because camouflaged targets can blend into their surroundings. In complex natural environments, background colors, textures, and lighting become intricate, increasing the difficulty of the detection task. In dynamic scenes, the appearance of camouflaged targets may change over time, posing a challenge to fixed parameters, while dynamic backgrounds can also introduce errors into the results. Changes in viewpoint and partial occlusion of the target can both affect the target's visibility, and traditional methods may not adequately respond to these changes, thus reducing detection performance. Therefore, it is essential to provide a method and system that can overcome the above problems.

[0005] The current common method for detecting camouflaged targets using drones and unmanned vehicles (UAVs) is to first have the UAV conduct initial target reconnaissance and localization, and then the UAV initiates and moves towards the target point. However, this method suffers from real-time issues in information transmission and processing. Since the UAV only begins its operation after receiving the information collected by the UAV, there is a certain time delay, preventing real-time information sharing and rapid response. One possible scenario is that after the UAV starts, the camouflaged target may have moved or escaped the location marked by the UAV while the UAV is en route, resulting in spatial asynchrony of target information and making it difficult for the UAV to locate and track the camouflaged target. For example, a method and related device for collaborative operation between an UAV and a swarm of UAVs (Patent Publication No. 115729247A, Publication Date 2023.03.03) proposes using a UAV controller to control the UAV to first photograph the surrounding road conditions, and then transmit the images to the image processing module of the UAV controller for processing. However, this method suffers from a relatively long delay; the UAV may not be stable enough during information collection, and the captured images may be unclear, leading to large image processing errors and making it difficult for the UAV to accurately obtain the target location. One operational method of a vehicle-mounted drone-assisted driving system (patent publication number 105825713B, publication date 2018.07.24) involves controlling a drone to photograph the surrounding road conditions via a drone controller and transmitting the images to an unmanned vehicle for processing. However, this method may face latency issues, resulting in insufficient real-time information transmission, and potential problems with the stability of the drone during information acquisition and the clarity of the captured images, thereby affecting the unmanned vehicle's ability to accurately perceive the target location. Summary of the Invention

[0006] This invention aims to solve the problem of inaccurate target location perception caused by insufficient real-time information transmission and large image processing errors, and provides a method and system for collaborative identification of camouflaged targets by drones and unmanned vehicles.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a method for collaborative detection of camouflaged targets by unmanned aerial vehicles and unmanned vehicles, comprising the following steps:

[0008] Step 1: Ensure the relative positions of the drone and unmanned aerial vehicle (UAV) remain consistent through the Global Positioning System (GPS) to achieve synchronized detection;

[0009] Step 2: The drone and the unmanned vehicle are equipped with infrared cameras, binocular cameras and radar sensors respectively to collect data on the target area;

[0010] Step 3: Process and analyze the collected data, and identify potential camouflaged targets through preprocessing and target detection algorithms;

[0011] Step 4: Employ pattern matching and information fusion technologies to improve the accuracy of the identification results, and issue corresponding warnings after confirming the disguised target.

[0012] In step 3 of this invention, the target detection and recognition process of the UAV and unmanned vehicle for the camouflaged target is specifically implemented as follows:

[0013] Step 3-1: The drone and unmanned vehicle are equipped with infrared cameras, binocular cameras, and radar sensors to simultaneously collect data on the surrounding environment. The infrared camera captures thermal images of the target, the binocular camera acquires images for 3D visual reconstruction, and the radar obtains information such as the target's position, speed, and shape.

[0014] Step 3-2: Preprocess the collected data, including data denoising, data augmentation, and data correction, to eliminate noise, improve image quality, and ensure data consistency;

[0015] Step 3-3: The preprocessed data is processed using an improved SSD algorithm for image processing and target detection. In the target detection stage, data from infrared cameras and binocular cameras are primarily analyzed to identify potential targets.

[0016] Steps 3-4: Once a potential target is identified, target recognition is performed. Target recognition mainly utilizes machine learning and pattern recognition algorithms to compare the target with a pre-trained model to determine the target's category. For identifying camouflaged targets, a training set containing camouflaged target samples can be built, and supervised learning can be performed.

[0017] Steps 3-5: By comparing and matching with the trained model, the target is classified and it is determined whether it is a camouflaged target. The determination criteria can be defined according to the characteristics of the camouflaged target, such as special texture, shape, or spatial distribution features.

[0018] Steps 3-6: Once a disguised target is identified, the system will output the results and issue an alert.

[0019] In step 4 of this invention, the matching of UAV and unmanned vehicle modes, the information fusion of various sensors aims to improve the accuracy and reliability of target detection; the specific implementation steps are as follows:

[0020] Step 4-1: Perform data time and spatial alignment to ensure that the data acquisition time points are consistent and that the spatial alignment is achieved by mapping data acquired by different sensors to the same coordinate system;

[0021] Sensors on unmanned aerial vehicles (UAVs) and unmanned vehicles (UGVs) may capture data at different times, which can cause data synchronization problems due to delays in signal transmission and processing (denoted as Δt). We synchronize the timestamps of the sensor data capture; the formula for time synchronization is:

[0022] t senor_uav =t senor_ugv +Δt

[0023] Where t senor_uav and t senor_ugv Δt represents the timestamps of the data collected by the sensors on the drone and the unmanned vehicle, respectively, and Δt represents the time difference caused by signal transmission and processing.

[0024] Spatial alignment is the process of ensuring that data captured by sensors carried by drones and autonomous vehicles can be mapped to the same coordinate system. This requires adjustments based on the sensor's position relative to its carrier platform. The global position vector can be decomposed into the carrier platform's position vector vec(p). xxx And the position vec(r) of the sensor relative to its platform. xxx_to_sensor Two parts.

[0025] Ultimately, the global position of the sensor can be expressed as the position of the platform plus its relative position to the sensor; the formula for spatial synchronization is:

[0026] vec(p) sensor_uav =vec(p) uav +vec(r) uav_to_sensor

[0027] vec(p) sensor_ugv =vec(p) ugv +vec(r) ugv_to_sensor ,

[0028] Where, vec(p) uav and vec(p) ugv vec(p) represents the position of the drone and unmanned vehicle in the global coordinate system. sensor_uav and vec(p) sensor_ugv vec(r) represents the position of sensors on drones and unmanned vehicles in the global coordinate system. uav_to_sensor and vec(r) ugv_to_sensor This represents the relative position vector from the reference point of the drone and the unmanned vehicle to their respective sensors;

[0029] Step 4-2: Once the data is aligned, we can use the SE attention module for feature optimization. This is applied to the three feature planes as follows:

[0030] Squeeze Global Average Pooling: Performs global average pooling on the feature descriptors of each channel, compressing global spatial information to the channel level; this is actually calculating the average of all elements for each channel, as shown in the formula below:

[0031]

[0032] Excitation is then applied, followed by mapping the trained weights W1 and W2 through a two-layer fully connected network. Each element in the output vector represents the importance weight of the corresponding channel; the formula is as follows:

[0033] A c =F ex (Z c ,W)=σ(W2δ(W1Z c ))

[0034] Scale Feature Recalibration: Finally, each channel in the original feature map is multiplied by its corresponding activation weight for feature recalibration; this strengthens important features and suppresses unimportant features; the formula is as follows:

[0035]

[0036] In the formula, U C Let U represent the c-th two-dimensional matrix, where the subscript c indicates the corresponding channel, and H and W represent the height and width of each feature map, respectively. c W1Z represents the weights indicating the importance of each channel. c W2δ represents the first fully connected operation, δ() represents the ReLU activation function, W2δ() represents the second fully connected operation, and σ() represents the sigmoid function. This indicates the result of feature recalibration;

[0037] The steps for applying the SE module to the acquired feature map F are as follows: First, apply the Squeeze operation to F to obtain compressed channel descriptors. Next, use the Excitation operation to obtain the weight coefficients for each channel. Finally, recalibrate the initial feature map F using these coefficients. The resulting feature map (\tilde{F}) will contain enhanced important channels and suppressed unimportant channels, which helps the neural network focus more on features that are more important for classifying or detecting camouflaged targets.

[0038] The advantage of SE attention fusion lies in its ability to efficiently process key information by weighting important features, effectively improving model performance. Its modules facilitate learning and modeling dependencies between channels, enhancing useful signals and suppressing irrelevant noise, further increasing the model's accuracy and robustness. This simple structure and low parameter count make it easily integrateable into various existing convolutional neural network architectures, offering broad applicability. Furthermore, SE attention fusion can adaptively adjust multi-layer feature responses, enhancing the network's ability to represent non-linear input data, thereby enriching the representation of information and optimizing model performance.

[0039] Step 4-3: After obtaining the optimized features from each sensor, these features are fused. During data fusion, establishing the correspondence between data from different sensors is also crucial. By establishing and updating a target tracking mechanism, data from different sensors are correlated to ensure the consistency and accuracy of the relevant information.

[0040] Step 4-4: Perform 3D reconstruction using image data from the binocular camera to obtain more accurate target location and shape information. By calculating the parallax and depth information between the binocular images, a 3D point cloud is constructed, further providing richer target features and shape descriptions.

[0041] Steps 4-5: During the fusion process, the credibility of the information provided by different sensors is assessed. By analyzing the accuracy, stability, and reliability of the sensors, the contributions of different sensors are weighed, and this information serves as the basis for decision-making.

[0042] This invention also provides a system for the collaborative detection of camouflaged targets by UAVs and unmanned vehicles, comprising four modules: positioning system calibration, data acquisition equipment configuration, data processing and analysis, and application of pattern matching and information fusion technologies. The system ensures synchronized operation of the UAV and unmanned vehicle through a global positioning system, and utilizes their onboard infrared cameras, binocular cameras, and radar sensors to efficiently acquire data from the target area. The collected data undergoes preprocessing and analysis, and based on this, an improved SSD algorithm and pattern recognition algorithm are used to identify potential camouflaged targets. Pattern matching and information fusion are then performed to improve the accuracy of the identification results.

[0043] This invention improves upon the SSD network as a benchmark for target detection and recognition of camouflaged targets in UAVs and unmanned vehicles. The SSD network detection framework is a classic single-stage convolutional neural network. Unlike two-stage target detection networks that first extract candidate regions and then perform classification detection and bounding box regression localization on the candidate regions, it directly uses regression to achieve target detection, which has a significant advantage in detection speed.

[0044] The improved SSD network, building upon the data-level fusion input from the previous section, incorporates feature-level fusion of visible light, infrared, and radar images. Based on the concept of domain adaptation in transfer learning, the Maximum Mean Difference (MMD) metric, which measures the distributional gap between different domains, is introduced into the SSD network. This MMD metric is used to calculate the feature gap between different domains at specific convolutional layers. By minimizing the distributional difference between different domains, the source domain data is fully utilized to improve the detection performance of the target domain.

[0045] The SE attention module in step 4-2 of this invention enhances the feature representations generated by the CNN by integrating an attention learning mechanism into the network, thereby helping to capture the spatial correlations between features. Through this mechanism, the network can learn to use global information more effectively, selectively focusing on useful information features and suppressing unwanted features.

[0046] Beneficial effects:

[0047] The method for collaborative identification of camouflaged targets by UAVs and unmanned vehicles in this invention utilizes multiple sensors, including infrared cameras, binocular cameras, and radar, for data acquisition and fusion analysis, significantly improving the accuracy and robustness of camouflaged target detection. GPS synchronous monitoring ensures spatiotemporal consistency, while the inclusion of an SE attention mechanism enhances feature representation and spatial correlation capture, further improving target detection performance. The 3D reconstruction technology of the binocular camera supports accurate acquisition of target morphology information, while data association and target tracking mechanisms maintain information consistency. Furthermore, the introduction of an improved SSD network and domain adaptation techniques (such as MMD) enhances the system's adaptability to diverse environments and the model's generalization ability. Ultimately, this collaborative approach not only increases the monitoring range and depth but also makes the system highly valuable for applications in various fields such as military, intelligence gathering, and urban planning. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the entire process of the method of the present invention.

[0049] Figure 2 This is a flowchart of the collaborative detection process between drones and unmanned vehicles.

[0050] Figure 3 This is a flowchart of pattern matching.

[0051] Figure 4 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0052] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0053] like Figure 1As shown, this invention comprises four modules: positioning system calibration, data acquisition equipment configuration, data processing and analysis, and application of pattern matching and information fusion technologies. This patent provides a method for collaboratively detecting camouflaged targets using unmanned aerial vehicles (UAVs) and unmanned vehicles. By utilizing infrared cameras, binocular cameras, and radar sensors, and through the fusion and collaboration of sensor data, the accuracy and robustness of target detection are effectively improved, demonstrating broad application potential. The specific implementation steps are as follows:

[0054] Step 1: Ensure the relative positions of the drone and unmanned aerial vehicle (UAV) remain consistent through the Global Positioning System (GPS) to achieve synchronized detection;

[0055] Step 2: The drone and the unmanned vehicle are equipped with infrared cameras, binocular cameras and radar sensors respectively to collect data on the target area;

[0056] Step 3: Process and analyze the collected data, and identify potential camouflaged targets through preprocessing and target detection algorithms;

[0057] Step 4: Employ pattern matching and information fusion technologies to improve the accuracy of the identification results, and issue corresponding warnings after confirming the disguised target.

[0058] like Figure 2 The flowchart of collaborative detection between drones and unmanned vehicles shows that in the target detection and recognition module, selecting appropriate algorithms, providing sufficient training samples, and adjusting recognition rules all contribute to improving the accuracy and robustness of camouflaged target recognition. Simultaneously, data fusion and collaborative work also play a crucial role in the target detection and recognition process, providing more accurate results and auxiliary judgments.

[0059] In the detection and identification process of camouflaged targets, infrared cameras, binocular cameras, and radar sensors equipped on drones and unmanned vehicles are used to collect real-time data on the surrounding environment. Infrared cameras are responsible for capturing thermal images of the target, binocular cameras are used to acquire images for 3D visual reconstruction, and radar acquires information such as the target's position, speed, and shape.

[0060] The next step is data preprocessing, which includes data denoising, data augmentation, and data correction. Data denoising typically uses filtering techniques to remove irrelevant noise from the data, while data augmentation expands the diversity of the data through image rotation, scaling, and inversion to improve the model's generalization ability. Data correction aims to eliminate data biases caused by equipment errors and the usage environment, making the data more accurately reflect the real situation.

[0061] Subsequently, the preprocessed data is processed for image processing and target detection using an improved SSD algorithm. This step involves analyzing data acquired by infrared and binocular cameras to determine the presence of potential targets.

[0062] Once a potential target is identified, target recognition is performed. This part mainly utilizes pre-trained machine learning and pattern recognition algorithms to compare the target to be identified with samples in the model to confirm the target's category. For the identification of camouflaged targets, we perform supervised learning on a training set containing camouflaged target samples.

[0063] Next, these target features are compared and matched with the trained model, allowing the system to classify and determine whether a target is a camouflaged target. The criteria for judging camouflaged targets may include features such as special textures, shapes, or spatial distributions that the target may possess.

[0064] Finally, once the camouflaged target is identified, the system outputs the results and issues warnings as needed. The key to the efficiency and accuracy of the entire process lies in the precise processing of data, which requires advanced algorithms and sufficient and diverse samples for training to enhance the system's robustness.

[0065] like Figure 3 As shown in the flowchart, the pattern matching process comprises five main modules: data alignment, data fusion, information association, target reconstruction, and credibility assessment. Developing a rigorous fusion strategy and association algorithm applicable to all scenarios is crucial. In this process, the SE attention module is introduced to enhance the feature representations generated by the convolutional neural network. It integrates an attention learning mechanism, effectively assisting in capturing the spatial correlation between features. The SE attention module works by guiding the network to focus more on valuable features and downplay unimportant ones through learned global information. It first performs channel squeezing, reducing the spatial information of each channel through global average pooling, leaving only global information describing the importance of features within each channel. Then, it performs activation operations, selectively highlighting or suppressing feature maps by learning and applying a set of nonlinear transformations. In this process, important features are highlighted, and unimportant information is suppressed, effectively enhancing the accuracy and robustness of feature representation.

[0066] The first step is data alignment, which addresses the challenges arising from the different data collection methods and characteristics of various sensors, such as infrared cameras, stereo cameras, and radar sensors. This process ensures that simultaneously acquired data is temporally aligned and spatially mapped to the same coordinate system.

[0067] After data alignment, the information collected by different sensors is fused, which requires the SE attention-based fusion network mentioned above.

[0068] In the data fusion process, linking the data collected by each sensor is also a crucial step. By establishing and updating the target tracking mechanism, the accuracy and consistency of the data collected from each sensor can be ensured.

[0069] Then, 3D reconstruction is performed using image data provided by the binocular cameras to obtain more accurate target location and shape information. By utilizing the parallax and depth information from the binocular images, richer target features and shape descriptions can be generated.

[0070] During the fusion process, a credibility assessment is also conducted, which involves analyzing the accuracy, stability, and reliability of the sensors to weigh the importance of each sensor's contribution to the overall system, thereby providing a basis for decision-making.

[0071] exist Figure 4 In the system architecture shown, the workbench, drone, and unmanned vehicle form a collaborative whole. Both the drone and the unmanned vehicle are equipped with high-precision binocular cameras, infrared cameras, and radar sensors to ensure all-around target perception capabilities. Seamless communication is achieved through the Global Positioning System (GPS) to coordinate the positioning information and target tracking activities of the drone and the unmanned vehicle. When the drone detects a camouflaged target using infrared thermal imaging technology, it activates its binocular cameras for 3D visual reconstruction. This reconstruction process not only enhances the visual representation of target features but also assists in the accurate analysis of the target's shape. Subsequently, the radar system locks onto the target's position, speed, and shape, and transmits the acquired target information to the workbench. After receiving the target data, the workbench performs coordinate transformation and further processes this information to specify the action commands for the unmanned vehicle.

[0072] The workstation transmits the converted coordinates of the camouflaged target. The autonomous vehicle (RV) adjusts its onboard sensors based on the received coordinates to obtain more precise target information. Finally, the RV feeds back the high-precision perception information to the workstation, which then optimizes its algorithms and autonomously executes subsequent task instructions. This entire process demonstrates an innovative information processing and decision-making workflow, improving the real-time performance and accuracy of camouflaged target detection systems and bringing groundbreaking technological advancements to the fields of security and surveillance.

[0073] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for collaborative detection of camouflaged targets using unmanned aerial vehicles (UAVs) and unmanned vehicles (UAVs), characterized in that, Includes the following steps: Step 1: Ensure the relative positions of the drone and the unmanned vehicle are kept consistent through the Global Positioning System (GPS) to achieve synchronous detection; Step 2: The drone and the unmanned vehicle are equipped with infrared cameras, binocular cameras and radar sensors respectively to collect data on the target area; Step 3: Process and analyze the collected data, and identify potential camouflaged targets through preprocessing and the improved SSD algorithm; The improved SSD network, based on data-level fusion input, sets up feature-level fusion of visible light and infrared images as well as radar images; The maximum mean difference metric is used to calculate the feature gap between different domains at a specific convolutional layer to improve object detection performance. Step 4: Employ pattern matching and information fusion technologies to improve the accuracy of the identification results, and issue corresponding warnings after confirming the disguised target; Specifically: Step 4-1: Perform data time and spatial alignment to ensure that the data acquisition time points are consistent and that the spatial alignment is achieved by mapping data acquired by different sensors to the same coordinate system; Step 4-2: Optimize the features of the aligned data using a data fusion network controlled by the SE attention module; Step 4-3: Establish data relationships and maintain information consistency through target tracking mechanisms; Step 4-4: Perform 3D reconstruction using binocular camera images to obtain a more accurate target shape; Steps 4-5: Assess the reliability of the sensor information and use it as a basis for decision-making.

2. The method for collaborative detection of camouflaged targets by unmanned aerial vehicles and unmanned vehicles according to claim 1, characterized in that, Step 3 specifically involves: Step 3-1: Use the sensors of drones and unmanned vehicles to collect data on the surrounding environment; Step 3-2: Preprocess the acquired data, including data denoising, data augmentation, and data correction operations, in order to eliminate noise, improve image quality, and ensure data consistency; Step 3-3: Image processing and target detection are performed using an improved SSD algorithm, and the preprocessed data is analyzed to identify potential targets; Steps 3-4: Using machine learning and pattern recognition algorithms, compare the target with the training set model to confirm the target category; Steps 3-5: After comparison and confirmation, classify the targets based on their characteristics and determine whether they are disguised targets; Steps 3-6: After confirming the disguised target, the system outputs the results and issues an alert.

3. The method for collaborative detection of camouflaged targets by unmanned aerial vehicles and unmanned vehicles according to claim 1, characterized in that, Step 4 is as follows: Step 4-1: Perform data time and spatial alignment to ensure that the data acquisition time points are consistent and that the spatial alignment is achieved by mapping data acquired by different sensors to the same coordinate system; Sensors on UAVs and UGVs may capture data at different times, which can cause data synchronization problems due to signal transmission and processing delays ∆t. Therefore, timestamps of the sensor data captures are synchronized. The formula for time synchronization is: in and These represent the timestamps of the data collected by sensors on the drone and the unmanned vehicle, respectively. This indicates the time difference caused by signal transmission and processing; Spatial alignment needs to be adjusted based on the sensor's position relative to its mounting platform; the global position vector can be decomposed into the position vector of the mounting platform. and the position of the sensor relative to its platform Ultimately, the global position of the sensor can be expressed as the position of the platform plus its relative position to the sensor; the formula for spatial synchronization is: , in, and This indicates the position of the drone and unmanned vehicle in the global coordinate system. and This indicates the position of the sensors on drones and unmanned vehicles in the global coordinate system. and This represents the relative position vector from the reference point of the drone and the unmanned vehicle to their respective sensors; Step 4-2: Optimize the features of the aligned data using a data fusion network controlled by the SE attention module. Squeez Global Average Pooling: Performs global average pooling on the feature descriptors of each channel, compressing global spatial information to the channel level. This is actually calculating the average of all elements for each channel, as shown in the formula below: Excitation activation: then using the trained weights and It is mapped using a two-layer fully connected network, and each element in the output vector represents the importance weight of the corresponding channel; the formula is as follows: Scale Feature Recalibration: Finally, each channel in the original feature map is multiplied by its corresponding activation weight for feature recalibration; this strengthens important features and suppresses unimportant features; the formula is as follows: In the formula, Let c represent the c-th two-dimensional matrix in U, where the subscript c indicates the corresponding channel, and H and W represent the height and width of each feature map, respectively. The weights representing the importance of each channel This indicates the first full-connect operation. Represents the ReLU activation function. This indicates the second fully connected operation. This represents the sigmoid function. This indicates the result of feature recalibration; Step 4-3: Establish data relationships and maintain information consistency through target tracking mechanisms; Step 4-4: Perform 3D reconstruction using binocular camera images to obtain a more accurate target shape; Steps 4-5: Assess the reliability of the sensor information and use it as a basis for decision-making.

4. The method for collaborative detection of camouflaged targets by unmanned aerial vehicles and unmanned vehicles according to claim 3, characterized in that, The steps for applying the SE attention module to the acquired feature map are as follows: first, apply the Squeeze operation to the feature map to obtain compressed channel descriptors; then, use the Excitation operation to obtain the weight coefficients of each channel; finally, use these coefficients to recalibrate the initial feature map; the processed feature map contains enhanced important channels and suppressed unimportant channels.

5. The method for collaborative detection of camouflaged targets by a drone and an unmanned vehicle according to claim 3, characterized in that, The SE attention module in step 4-2 enhances the feature representations generated by the CNN by integrating an attention learning mechanism into the network, thereby helping to capture the spatial correlation between features, and selectively focusing on useful information features while suppressing unwanted features.

6. A system for collaborative detection of camouflaged targets using both unmanned aerial vehicles (UAVs) and unmanned vehicles (UAVs), characterized in that, The method described in any one of claims 1-5 ensures the synchronous operation of UAVs and unmanned vehicles through a global positioning system, and utilizes the infrared cameras, binocular cameras and radar sensors carried by both to efficiently collect data on the target area; wherein, the collected data is preprocessed and analyzed, and on this basis, an improved SSD algorithm and pattern recognition algorithm are used to identify potential camouflaged targets, and pattern matching and information fusion are performed to improve the accuracy of the identification results.

Images